Improved robustness of noisy speech HMMs based on weighted variance expansion

The spectrum of noise and SNR often vary abruptly, due to the non-stationary noise under field conditions. The performance of speech recognition degrades rapidly when the noise conditions in the recognition process are different from those in the process of training or adaptation; therefore, it is necessary to make HMMs robust to abrupt variations of noise. We propose a method to modify the output probability at the state sensitive to noise by using a weighted variance expansion based on the power of state or probability distribution, in order to improve the performance. The effectiveness of this method was examined in two types of noisy speech HMMs (one was trained with a specific SNR, the other was trained with five kinds of SNRs), through the evaluation experiments of speaker independent word recognition using the noise of two factories. As a result, this method improved the robustness of the HMMs against the variation of noise conditions (noise type and SNR).